# 50 GPT-6.1 Sol agents build a Minecraft civilization with leaders, rules and rationing

> Y Combinator Labs and Olam Labs put 50 agents in a Minecraft world, where they appointed a leader, wrote rules and built a food-rationing ledger.

- Status: reported (One outlet, not independently verified yet.)
- Topic: Research
- Published: 2026-11-04T13:13:54+00:00
- Updated: 2026-11-04T13:13:54+00:00
- Web page: https://ai-news.ikangai.com/story/50-gpt-6-1-sol-agents-build-a-minecraft-civilization-with-leaders-rules-and-rati/
- Seen in: Forward Future (Matthew Berman)

Y Combinator Labs and Olam Labs put 50 GPT-6.1 Sol agents into a survival Minecraft world for three hours with one goal: build a thriving civilization. The agents sent about 5,900 messages, appointed a leader, wrote rules and permits, and built a food-rationing ledger. The labs report that 22 of the run's 34 agent deaths occurred during the first night, and those losses shaped the rest of the run.

The team had no existing harness for the run, so it built one and open-sourced it on GitHub. The harness uses Mineflayer, a library for scripting Minecraft players, with the Pi agent loop handling the model session. Agents could observe the world, remember things, take actions, send messages and write scripts. They also shared an agent directory and a shared sandbox and could send direct messages. The repository's default model is GPT-6.1 Terra, while the experiment used Sol.

The early hours were a rush for wood. The agents harvested almost all trees within reach of the spawn point in 90 seconds. A hierarchy formed quickly and followed the order of the agent directory. Ada, listed first, became the de facto leader. Agents asked her for approvals, and she wrote a settlement plan in the shared sandbox. The plan assigned storage chests, farming, stone mining and a shelter. It also told agents not to harvest immature wheat and to preserve saplings, because wood had become scarce.

The first night was deadly. Rafe was mining coal when a zombie chased him. He fled, but the zombie caught him. Vera, lighting a quarry with torches, killed a zombie and a skeleton, then reported the area was overrun by spiders, skeletons and creepers. A spider killed her. Death carried no penalty beyond the loss of accumulated gear. An agent that died respawned, and a short line was added to a death_log.md file. The labs say the first night's deaths permanently increased the agents' caution, and many ambitious early tasks were abandoned.

The caution persisted. After the third night, Ada wrote that agents should stop optional bulk excavation, building and distant scouting. She told them to "Idle in lit closed shelters until assigned short productive tasks." The shared plans then shifted from settlement to survival. A food-operations memo covered fishing at a protected riverside station, bonemeal from a composter, and bread made from combined wheat.

The rationing system ran through Ines. She took requests, tracked what went into and out of the chests, and kept a file called food-queue.md on the shared machine. The file labels its numbers as reported values, not live readings. It also sets out ration math. Cooked beef or pork restored 8 food, cooked chicken or mutton 6, bread 5 and an apple 4, against a target of 18 that the file treats as the level for regeneration.

Ines's rule was to confirm a recipient's inventory and eating before marking food as served, because a pickup alone proved little. The ledger shows the difficulty. Noor appears as unserved across many updates. Two cooked chickens from Rafe were distributed by Ada. Basil was confirmed to have eaten one. Esme picked up and ate a chicken intended for Noor, and the file later marked the misdelivery as resolved while Noor remained unserved. Separately, Esme picked up uneaten mutton and bread meant for Tove and Nikhil, which were deposited in an indoor chest, and Ines reserved each item by direct message.

This is one game experiment, not a model of human society.

## Why it matters

Multi-agent systems can quickly form hierarchies and cautious behavior, which matters for teams designing agent swarms, though the results come from a game setting.

## Original source

- [GitHub - yc-software/minecraft-agent-harness: Run language-model agents in Minecraft with Mineflayer tools, persistent sessions and shell sandboxes. · GitHub](https://github.com/yc-software/minecraft-agent-harness)

## Further reporting

- [Agents in Games · Labs](https://labs.ycombinator.com/projects/games/)

---

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it.